Investment Realm has learned that the world model company FaceMind recently completed a Pre-A round financing of tens of millions of yuan. The investors were StarLink Capital, with existing shareholder 360 significantly over-subscribing its follow-on investment.
It is reported that FaceMind's new round of financing is already in progress, with financial advisors such as Shendu Capital acting as FA. Investment institutions have already expressed investment intentions.
This is a young AI company. At the helm is Lu Hongyuan, born after 1995, who founded FaceMind while still a student. Over the past two years, the company has pivoted from focusing on on-device full-modal models to more fundamental world models.
As AI enters screens, software, and robots, understanding the world is becoming the next major topic.
Post-95s Doctoral Graduate Leads the Team
A World Model Team Emerges
The story of FaceMind begins with Lu Hongyuan.
A post-95s founder, Lu Hongyuan completed his bachelor's and master's studies at Imperial College London and earned his PhD from the Natural Language Processing Laboratory at The Chinese University of Hong Kong, under the supervision of Professor Lin Wei. He has long researched natural language processing and the underlying mechanisms of large models. During his PhD, he produced 14 first-author/corresponding-author papers at top-tier conferences, with several becoming highly cited in the field.
FaceMind was founded in 2023, initially targeting the research, development, and application of on-device full-modal models.
What truly brought them to public attention was the earlier discussion about "Ma Jiaqi causing a large model to fail." At that time, a large model could accurately state Ma Jiaqi's relevant career details but couldn't stably output the three characters for "Ma Jiaqi." An ordinary name unexpectedly exposed a fundamental issue in large models' language processing: before text enters a model, it is first tokenized; when the model encounters low-frequency words, uncommon names, or words from less common languages, comprehension and generation can become unstable.
Lu Hongyuan's team focused on this issue earlier. In 2025, they published a paper related to SLoW, discussing how low-frequency words affect the translation performance of large models; by 2026, their paper on Adam's Law further pushed the issue to the sentence level—the more frequent and common the expression for the same meaning, the more easily it tends to be processed and learned by the model.
More unexpectedly, the technology related to this paper was adopted by Anthropic and was liked and reposted by an Anthropic investor on X. A Chinese post-95s researcher's judgment on the underlying principles of large models was thus seen by more people.
Following this line further, FaceMind began shifting its focus to world models.
Simply put, while large language models are good at predicting the next piece of text, world models aim to predict what will happen next in an environment. In the context of screens, it's about GUI Agents understanding web pages, documents, buttons, and user intent; in the field of robotics, it's about understanding space, actions, and task outcomes.
FaceMind's self-developed world model system is developed precisely around this direction. The company aims to enhance the model's stability in long-term temporal prediction, screen understanding, and embodied tasks through a cyclic, iterative, and parameter-efficient model architecture.
DieDieShe (叠叠社) serves as an early testing ground for this capability. On the surface, it appears to be an AI bullet-comment (danmu) product that can generate interactive bullet comments in real-time based on the web pages, documents, videos, or game content a user is browsing. Looking deeper, for a GUI Agent to complete tasks, it must understand the screen, comprehend page structure, judge button positions, and predict results after clicks. Each page jump, input feedback, and task completion forms a type of high-density world model data.
This is also the opportunity FaceMind wants to seize: world models are becoming a new underlying entry point for AI.
StarLink Capital and 360 Invest
The Hottest Battlefield in Embodied AI
The latest financing comes to light.
Recently, FaceMind announced the completion of a tens-of-millions-yuan Pre-A round financing. This round not only introduced new investor StarLink Capital but also secured a significant oversubscription from existing shareholder 360.
Xiang Qiqi, the pre-investment head of 360 Group, stated, "Dr. Lu is one of the most outstanding young AI researchers I have ever met."
In his view, Lu Hongyuan focuses not on local optimizations but on the fundamental principles and architectural innovations of models. While the industry was still discussing the concept of world models, FaceMind had already started training world models from scratch, achieving industry SOTA-level results in various benchmarks. Subsequently, Adam's Law gained attention and validation from overseas leading model developer Anthropic, and the team's latest proposed Loop cyclic architecture further explores long-term temporal training issues in world models.
"The iteration speed is astonishing. Before every communication, I first look at their latest published papers and technical reports," Xiang Qiqi remarked, truly experiencing what it means to "invest once, learn for a lifetime."
Li Wenjue, a partner at StarLink Capital, said the most prominent characteristic of the FaceMind team is its combination of solid research capabilities and complex engineering implementation abilities. The core team members have long been deeply involved in fundamental AI technology, capable of forming independent judgments on cutting-edge directions and rapidly validating research results in real scenarios.
"What we see is a team with high talent density, forward-looking technical judgment, and strong execution capabilities." In her view, Lu Hongyuan embodies both the exploratory drive of a young researcher and the action-oriented mindset of an entrepreneur, capable of leading the team to continually tackle high-difficulty problems and translate technical judgments into clear R&D directions. This founder trait and team cohesion were important reasons for StarLink Capital's decision to invest.
Over the past year, world models have become a new keyword in the AI industry. Amidst the excitement, disagreements have also emerged: Will the next stage of competition continue to rely on larger data and parameters, or will it improve model efficiency in utilizing limited data through new architectures?
FaceMind has chosen the latter.
According to the introduction, the core characteristics of the company's self-developed model are cyclic iteration and parameter efficiency. Simply put, it attempts to enable the model, under the same parameter scale, to acquire stronger long-term temporal prediction and environmental reasoning capabilities. The company disclosed that its 1B-scale model's performance already rivals that of internationally comparable strong models, with improved parameter efficiency.
Currently, FaceMind has begun validating this model capability in multiple scenarios. Information shows its world model capability has been validated in simulated embodied environments, GUI Agent environments, and real-world robotic arm environments. For downstream applications, the company plans to provide a full suite of capabilities—from scenario validation, model training, and architectural deployment to inference services and continuous optimization—for partners such as robot OEMs, content platforms, chip and cloud vendors.
In Lu Hongyuan's view, the opportunity for world models will unfold alongside GUI Agents and embodied intelligence. At that point, the competition among models will be about their ability to understand tasks, predict changes, and stably complete actions. After the financing round, FaceMind will continue to invest in world model R&D and multi-scenario validation.
A young company is squeezing its way onto the table for the next generation of AI infrastructure.
This article is from the WeChat public account "Investment Realm AI," author: Wang Lu
